1 // SPDX-License-Identifier: GPL-2.0
5 * Builtin report command: Analyze the perf.data input file,
6 * look up and read DSOs and symbol information and display
7 * a histogram of results, along various sorting keys.
11 #include "util/config.h"
13 #include "util/annotate.h"
14 #include "util/color.h"
16 #include <linux/list.h>
17 #include <linux/rbtree.h>
18 #include <linux/err.h>
19 #include <linux/zalloc.h>
21 #include "util/symbol.h"
22 #include "util/map_symbol.h"
23 #include "util/mem-events.h"
24 #include "util/branch.h"
25 #include "util/callchain.h"
26 #include "util/values.h"
29 #include "util/debug.h"
30 #include "util/evlist.h"
31 #include "util/evsel.h"
32 #include "util/evswitch.h"
33 #include "util/header.h"
34 #include "util/session.h"
35 #include "util/srcline.h"
36 #include "util/tool.h"
38 #include <subcmd/parse-options.h>
39 #include <subcmd/exec-cmd.h>
40 #include "util/parse-events.h"
42 #include "util/thread.h"
43 #include "util/sort.h"
44 #include "util/hist.h"
45 #include "util/data.h"
46 #include "arch/common.h"
47 #include "util/time-utils.h"
48 #include "util/auxtrace.h"
49 #include "util/units.h"
50 #include "util/branch.h"
51 #include "util/util.h" // perf_tip()
53 #include "ui/progress.h"
54 #include "util/block-info.h"
60 #include <linux/ctype.h>
62 #include <linux/bitmap.h>
63 #include <linux/string.h>
64 #include <linux/stringify.h>
65 #include <linux/time64.h>
66 #include <sys/types.h>
69 #include <linux/mman.h>
72 struct perf_tool tool;
73 struct perf_session *session;
74 struct evswitch evswitch;
75 bool use_tui, use_gtk, use_stdio;
78 bool inverted_callchain;
85 bool nonany_branch_mode;
89 struct perf_read_values show_threads_values;
90 struct annotation_options annotation_opts;
91 const char *pretty_printing_style;
93 const char *symbol_filter_str;
95 struct perf_time_interval *ptime_range;
103 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
104 struct branch_type_stat brtype_stat;
106 bool total_cycles_mode;
107 struct block_report *block_reports;
108 int nr_block_reports;
111 static int report__config(const char *var, const char *value, void *cb)
113 struct report *rep = cb;
115 if (!strcmp(var, "report.group")) {
116 symbol_conf.event_group = perf_config_bool(var, value);
119 if (!strcmp(var, "report.percent-limit")) {
120 double pcnt = strtof(value, NULL);
122 rep->min_percent = pcnt;
123 callchain_param.min_percent = pcnt;
126 if (!strcmp(var, "report.children")) {
127 symbol_conf.cumulate_callchain = perf_config_bool(var, value);
130 if (!strcmp(var, "report.queue-size"))
131 return perf_config_u64(&rep->queue_size, var, value);
133 if (!strcmp(var, "report.sort_order")) {
134 default_sort_order = strdup(value);
141 static int hist_iter__report_callback(struct hist_entry_iter *iter,
142 struct addr_location *al, bool single,
146 struct report *rep = arg;
147 struct hist_entry *he = iter->he;
148 struct evsel *evsel = iter->evsel;
149 struct perf_sample *sample = iter->sample;
151 struct branch_info *bi;
153 if (!ui__has_annotation() && !rep->symbol_ipc)
156 if (sort__mode == SORT_MODE__BRANCH) {
157 bi = he->branch_info;
158 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
162 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
164 } else if (rep->mem_mode) {
166 err = addr_map_symbol__inc_samples(&mi->daddr, sample, evsel);
170 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
172 } else if (symbol_conf.cumulate_callchain) {
174 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
176 err = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
183 static int hist_iter__branch_callback(struct hist_entry_iter *iter,
184 struct addr_location *al __maybe_unused,
185 bool single __maybe_unused,
188 struct hist_entry *he = iter->he;
189 struct report *rep = arg;
190 struct branch_info *bi = he->branch_info;
191 struct perf_sample *sample = iter->sample;
192 struct evsel *evsel = iter->evsel;
195 branch_type_count(&rep->brtype_stat, &bi->flags,
196 bi->from.addr, bi->to.addr);
198 if (!ui__has_annotation() && !rep->symbol_ipc)
201 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
205 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
211 static void setup_forced_leader(struct report *report,
212 struct evlist *evlist)
214 if (report->group_set)
215 perf_evlist__force_leader(evlist);
218 static int process_feature_event(struct perf_session *session,
219 union perf_event *event)
221 struct report *rep = container_of(session->tool, struct report, tool);
223 if (event->feat.feat_id < HEADER_LAST_FEATURE)
224 return perf_event__process_feature(session, event);
226 if (event->feat.feat_id != HEADER_LAST_FEATURE) {
227 pr_err("failed: wrong feature ID: %" PRI_lu64 "\n",
228 event->feat.feat_id);
233 * (feat_id = HEADER_LAST_FEATURE) is the end marker which
234 * means all features are received, now we can force the
237 setup_forced_leader(rep, session->evlist);
241 static int process_sample_event(struct perf_tool *tool,
242 union perf_event *event,
243 struct perf_sample *sample,
245 struct machine *machine)
247 struct report *rep = container_of(tool, struct report, tool);
248 struct addr_location al;
249 struct hist_entry_iter iter = {
252 .hide_unresolved = symbol_conf.hide_unresolved,
253 .add_entry_cb = hist_iter__report_callback,
257 if (perf_time__ranges_skip_sample(rep->ptime_range, rep->range_num,
262 if (evswitch__discard(&rep->evswitch, evsel))
265 if (machine__resolve(machine, &al, sample) < 0) {
266 pr_debug("problem processing %d event, skipping it.\n",
272 al.thread->lbr_stitch_enable = true;
274 if (symbol_conf.hide_unresolved && al.sym == NULL)
277 if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap))
280 if (sort__mode == SORT_MODE__BRANCH) {
282 * A non-synthesized event might not have a branch stack if
283 * branch stacks have been synthesized (using itrace options).
285 if (!sample->branch_stack)
288 iter.add_entry_cb = hist_iter__branch_callback;
289 iter.ops = &hist_iter_branch;
290 } else if (rep->mem_mode) {
291 iter.ops = &hist_iter_mem;
292 } else if (symbol_conf.cumulate_callchain) {
293 iter.ops = &hist_iter_cumulative;
295 iter.ops = &hist_iter_normal;
299 al.map->dso->hit = 1;
301 if (ui__has_annotation() || rep->symbol_ipc || rep->total_cycles_mode) {
302 hist__account_cycles(sample->branch_stack, &al, sample,
303 rep->nonany_branch_mode,
307 ret = hist_entry_iter__add(&iter, &al, rep->max_stack, rep);
309 pr_debug("problem adding hist entry, skipping event\n");
311 addr_location__put(&al);
315 static int process_read_event(struct perf_tool *tool,
316 union perf_event *event,
317 struct perf_sample *sample __maybe_unused,
319 struct machine *machine __maybe_unused)
321 struct report *rep = container_of(tool, struct report, tool);
323 if (rep->show_threads) {
324 const char *name = evsel__name(evsel);
325 int err = perf_read_values_add_value(&rep->show_threads_values,
326 event->read.pid, event->read.tid,
338 /* For pipe mode, sample_type is not currently set */
339 static int report__setup_sample_type(struct report *rep)
341 struct perf_session *session = rep->session;
342 u64 sample_type = perf_evlist__combined_sample_type(session->evlist);
343 bool is_pipe = perf_data__is_pipe(session->data);
345 if (session->itrace_synth_opts->callchain ||
346 session->itrace_synth_opts->add_callchain ||
348 perf_header__has_feat(&session->header, HEADER_AUXTRACE) &&
349 !session->itrace_synth_opts->set))
350 sample_type |= PERF_SAMPLE_CALLCHAIN;
352 if (session->itrace_synth_opts->last_branch ||
353 session->itrace_synth_opts->add_last_branch)
354 sample_type |= PERF_SAMPLE_BRANCH_STACK;
356 if (!is_pipe && !(sample_type & PERF_SAMPLE_CALLCHAIN)) {
357 if (perf_hpp_list.parent) {
358 ui__error("Selected --sort parent, but no "
359 "callchain data. Did you call "
360 "'perf record' without -g?\n");
363 if (symbol_conf.use_callchain &&
364 !symbol_conf.show_branchflag_count) {
365 ui__error("Selected -g or --branch-history.\n"
366 "But no callchain or branch data.\n"
367 "Did you call 'perf record' without -g or -b?\n");
370 } else if (!callchain_param.enabled &&
371 callchain_param.mode != CHAIN_NONE &&
372 !symbol_conf.use_callchain) {
373 symbol_conf.use_callchain = true;
374 if (callchain_register_param(&callchain_param) < 0) {
375 ui__error("Can't register callchain params.\n");
380 if (symbol_conf.cumulate_callchain) {
381 /* Silently ignore if callchain is missing */
382 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
383 symbol_conf.cumulate_callchain = false;
384 perf_hpp__cancel_cumulate();
388 if (sort__mode == SORT_MODE__BRANCH) {
390 !(sample_type & PERF_SAMPLE_BRANCH_STACK)) {
391 ui__error("Selected -b but no branch data. "
392 "Did you call perf record without -b?\n");
397 if (sort__mode == SORT_MODE__MEMORY) {
398 if (!is_pipe && !(sample_type & PERF_SAMPLE_DATA_SRC)) {
399 ui__error("Selected --mem-mode but no mem data. "
400 "Did you call perf record without -d?\n");
405 if (symbol_conf.use_callchain || symbol_conf.cumulate_callchain) {
406 if ((sample_type & PERF_SAMPLE_REGS_USER) &&
407 (sample_type & PERF_SAMPLE_STACK_USER)) {
408 callchain_param.record_mode = CALLCHAIN_DWARF;
409 dwarf_callchain_users = true;
410 } else if (sample_type & PERF_SAMPLE_BRANCH_STACK)
411 callchain_param.record_mode = CALLCHAIN_LBR;
413 callchain_param.record_mode = CALLCHAIN_FP;
416 if (rep->stitch_lbr && (callchain_param.record_mode != CALLCHAIN_LBR)) {
417 ui__warning("Can't find LBR callchain. Switch off --stitch-lbr.\n"
418 "Please apply --call-graph lbr when recording.\n");
419 rep->stitch_lbr = false;
422 /* ??? handle more cases than just ANY? */
423 if (!(perf_evlist__combined_branch_type(session->evlist) &
424 PERF_SAMPLE_BRANCH_ANY))
425 rep->nonany_branch_mode = true;
427 #if !defined(HAVE_LIBUNWIND_SUPPORT) && !defined(HAVE_DWARF_SUPPORT)
428 if (dwarf_callchain_users) {
429 ui__warning("Please install libunwind or libdw "
430 "development packages during the perf build.\n");
437 static void sig_handler(int sig __maybe_unused)
442 static size_t hists__fprintf_nr_sample_events(struct hists *hists, struct report *rep,
443 const char *evname, FILE *fp)
447 unsigned long nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
448 u64 nr_events = hists->stats.total_period;
449 struct evsel *evsel = hists_to_evsel(hists);
451 size_t size = sizeof(buf);
452 int socked_id = hists->socket_filter;
457 if (symbol_conf.filter_relative) {
458 nr_samples = hists->stats.nr_non_filtered_samples;
459 nr_events = hists->stats.total_non_filtered_period;
462 if (evsel__is_group_event(evsel)) {
465 evsel__group_desc(evsel, buf, size);
468 for_each_group_member(pos, evsel) {
469 const struct hists *pos_hists = evsel__hists(pos);
471 if (symbol_conf.filter_relative) {
472 nr_samples += pos_hists->stats.nr_non_filtered_samples;
473 nr_events += pos_hists->stats.total_non_filtered_period;
475 nr_samples += pos_hists->stats.nr_events[PERF_RECORD_SAMPLE];
476 nr_events += pos_hists->stats.total_period;
481 nr_samples = convert_unit(nr_samples, &unit);
482 ret = fprintf(fp, "# Samples: %lu%c", nr_samples, unit);
483 if (evname != NULL) {
484 ret += fprintf(fp, " of event%s '%s'",
485 evsel->core.nr_members > 1 ? "s" : "", evname);
489 ret += fprintf(fp, " (time slices: %s)", rep->time_str);
491 if (symbol_conf.show_ref_callgraph &&
492 strstr(evname, "call-graph=no")) {
493 ret += fprintf(fp, ", show reference callgraph");
497 ret += fprintf(fp, "\n# Total weight : %" PRIu64, nr_events);
498 ret += fprintf(fp, "\n# Sort order : %s", sort_order ? : default_mem_sort_order);
500 ret += fprintf(fp, "\n# Event count (approx.): %" PRIu64, nr_events);
503 ret += fprintf(fp, "\n# Processor Socket: %d", socked_id);
505 return ret + fprintf(fp, "\n#\n");
508 static int perf_evlist__tui_block_hists_browse(struct evlist *evlist,
514 evlist__for_each_entry(evlist, pos) {
515 ret = report__browse_block_hists(&rep->block_reports[i++].hist,
516 rep->min_percent, pos,
517 &rep->session->header.env,
518 &rep->annotation_opts);
526 static int perf_evlist__tty_browse_hists(struct evlist *evlist,
534 fprintf(stdout, "#\n# Total Lost Samples: %" PRIu64 "\n#\n",
535 evlist->stats.total_lost_samples);
538 evlist__for_each_entry(evlist, pos) {
539 struct hists *hists = evsel__hists(pos);
540 const char *evname = evsel__name(pos);
542 if (symbol_conf.event_group && !evsel__is_group_leader(pos))
545 hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
547 if (rep->total_cycles_mode) {
548 report__browse_block_hists(&rep->block_reports[i++].hist,
549 rep->min_percent, pos,
554 hists__fprintf(hists, !quiet, 0, 0, rep->min_percent, stdout,
555 !(symbol_conf.use_callchain ||
556 symbol_conf.show_branchflag_count));
557 fprintf(stdout, "\n\n");
561 fprintf(stdout, "#\n# (%s)\n#\n", help);
563 if (rep->show_threads) {
564 bool style = !strcmp(rep->pretty_printing_style, "raw");
565 perf_read_values_display(stdout, &rep->show_threads_values,
567 perf_read_values_destroy(&rep->show_threads_values);
570 if (sort__mode == SORT_MODE__BRANCH)
571 branch_type_stat_display(stdout, &rep->brtype_stat);
576 static void report__warn_kptr_restrict(const struct report *rep)
578 struct map *kernel_map = machine__kernel_map(&rep->session->machines.host);
579 struct kmap *kernel_kmap = kernel_map ? map__kmap(kernel_map) : NULL;
581 if (perf_evlist__exclude_kernel(rep->session->evlist))
584 if (kernel_map == NULL ||
585 (kernel_map->dso->hit &&
586 (kernel_kmap->ref_reloc_sym == NULL ||
587 kernel_kmap->ref_reloc_sym->addr == 0))) {
589 "As no suitable kallsyms nor vmlinux was found, kernel samples\n"
590 "can't be resolved.";
592 if (kernel_map && map__has_symbols(kernel_map)) {
593 desc = "If some relocation was applied (e.g. "
594 "kexec) symbols may be misresolved.";
598 "Kernel address maps (/proc/{kallsyms,modules}) were restricted.\n\n"
599 "Check /proc/sys/kernel/kptr_restrict before running 'perf record'.\n\n%s\n\n"
600 "Samples in kernel modules can't be resolved as well.\n\n",
605 static int report__gtk_browse_hists(struct report *rep, const char *help)
607 int (*hist_browser)(struct evlist *evlist, const char *help,
608 struct hist_browser_timer *timer, float min_pcnt);
610 hist_browser = dlsym(perf_gtk_handle, "perf_evlist__gtk_browse_hists");
612 if (hist_browser == NULL) {
613 ui__error("GTK browser not found!\n");
617 return hist_browser(rep->session->evlist, help, NULL, rep->min_percent);
620 static int report__browse_hists(struct report *rep)
623 struct perf_session *session = rep->session;
624 struct evlist *evlist = session->evlist;
625 const char *help = perf_tip(system_path(TIPDIR));
628 /* fallback for people who don't install perf ;-) */
629 help = perf_tip(DOCDIR);
631 help = "Cannot load tips.txt file, please install perf!";
634 switch (use_browser) {
636 if (rep->total_cycles_mode) {
637 ret = perf_evlist__tui_block_hists_browse(evlist, rep);
641 ret = perf_evlist__tui_browse_hists(evlist, help, NULL,
643 &session->header.env,
644 true, &rep->annotation_opts);
646 * Usually "ret" is the last pressed key, and we only
647 * care if the key notifies us to switch data file.
649 if (ret != K_SWITCH_INPUT_DATA && ret != K_RELOAD)
653 ret = report__gtk_browse_hists(rep, help);
656 ret = perf_evlist__tty_browse_hists(evlist, rep, help);
663 static int report__collapse_hists(struct report *rep)
665 struct ui_progress prog;
669 ui_progress__init(&prog, rep->nr_entries, "Merging related events...");
671 evlist__for_each_entry(rep->session->evlist, pos) {
672 struct hists *hists = evsel__hists(pos);
675 hists->symbol_filter_str = rep->symbol_filter_str;
677 hists->socket_filter = rep->socket_filter;
679 ret = hists__collapse_resort(hists, &prog);
683 /* Non-group events are considered as leader */
684 if (symbol_conf.event_group && !evsel__is_group_leader(pos)) {
685 struct hists *leader_hists = evsel__hists(pos->leader);
687 hists__match(leader_hists, hists);
688 hists__link(leader_hists, hists);
692 ui_progress__finish();
696 static int hists__resort_cb(struct hist_entry *he, void *arg)
698 struct report *rep = arg;
699 struct symbol *sym = he->ms.sym;
701 if (rep->symbol_ipc && sym && !sym->annotate2) {
702 struct evsel *evsel = hists_to_evsel(he->hists);
704 symbol__annotate2(&he->ms, evsel,
705 &annotation__default_options, NULL);
711 static void report__output_resort(struct report *rep)
713 struct ui_progress prog;
716 ui_progress__init(&prog, rep->nr_entries, "Sorting events for output...");
718 evlist__for_each_entry(rep->session->evlist, pos) {
719 perf_evsel__output_resort_cb(pos, &prog,
720 hists__resort_cb, rep);
723 ui_progress__finish();
726 static void stats_setup(struct report *rep)
728 memset(&rep->tool, 0, sizeof(rep->tool));
729 rep->tool.no_warn = true;
732 static int stats_print(struct report *rep)
734 struct perf_session *session = rep->session;
736 perf_session__fprintf_nr_events(session, stdout);
740 static void tasks_setup(struct report *rep)
742 memset(&rep->tool, 0, sizeof(rep->tool));
743 rep->tool.ordered_events = true;
744 if (rep->mmaps_mode) {
745 rep->tool.mmap = perf_event__process_mmap;
746 rep->tool.mmap2 = perf_event__process_mmap2;
748 rep->tool.comm = perf_event__process_comm;
749 rep->tool.exit = perf_event__process_exit;
750 rep->tool.fork = perf_event__process_fork;
751 rep->tool.no_warn = true;
755 struct thread *thread;
756 struct list_head list;
757 struct list_head children;
760 static struct task *tasks_list(struct task *task, struct machine *machine)
762 struct thread *parent_thread, *thread = task->thread;
763 struct task *parent_task;
765 /* Already listed. */
766 if (!list_empty(&task->list))
769 /* Last one in the chain. */
770 if (thread->ppid == -1)
773 parent_thread = machine__find_thread(machine, -1, thread->ppid);
775 return ERR_PTR(-ENOENT);
777 parent_task = thread__priv(parent_thread);
778 list_add_tail(&task->list, &parent_task->children);
779 return tasks_list(parent_task, machine);
782 static size_t maps__fprintf_task(struct maps *maps, int indent, FILE *fp)
787 maps__for_each_entry(maps, map) {
788 printed += fprintf(fp, "%*s %" PRIx64 "-%" PRIx64 " %c%c%c%c %08" PRIx64 " %" PRIu64 " %s\n",
789 indent, "", map->start, map->end,
790 map->prot & PROT_READ ? 'r' : '-',
791 map->prot & PROT_WRITE ? 'w' : '-',
792 map->prot & PROT_EXEC ? 'x' : '-',
793 map->flags & MAP_SHARED ? 's' : 'p',
795 map->dso->id.ino, map->dso->name);
801 static void task__print_level(struct task *task, FILE *fp, int level)
803 struct thread *thread = task->thread;
805 int comm_indent = fprintf(fp, " %8d %8d %8d |%*s",
806 thread->pid_, thread->tid, thread->ppid,
809 fprintf(fp, "%s\n", thread__comm_str(thread));
811 maps__fprintf_task(thread->maps, comm_indent, fp);
813 if (!list_empty(&task->children)) {
814 list_for_each_entry(child, &task->children, list)
815 task__print_level(child, fp, level + 1);
819 static int tasks_print(struct report *rep, FILE *fp)
821 struct perf_session *session = rep->session;
822 struct machine *machine = &session->machines.host;
823 struct task *tasks, *task;
824 unsigned int nr = 0, itask = 0, i;
829 * No locking needed while accessing machine->threads,
830 * because --tasks is single threaded command.
833 /* Count all the threads. */
834 for (i = 0; i < THREADS__TABLE_SIZE; i++)
835 nr += machine->threads[i].nr;
837 tasks = malloc(sizeof(*tasks) * nr);
841 for (i = 0; i < THREADS__TABLE_SIZE; i++) {
842 struct threads *threads = &machine->threads[i];
844 for (nd = rb_first_cached(&threads->entries); nd;
846 task = tasks + itask++;
848 task->thread = rb_entry(nd, struct thread, rb_node);
849 INIT_LIST_HEAD(&task->children);
850 INIT_LIST_HEAD(&task->list);
851 thread__set_priv(task->thread, task);
856 * Iterate every task down to the unprocessed parent
857 * and link all in task children list. Task with no
858 * parent is added into 'list'.
860 for (itask = 0; itask < nr; itask++) {
861 task = tasks + itask;
863 if (!list_empty(&task->list))
866 task = tasks_list(task, machine);
868 pr_err("Error: failed to process tasks\n");
870 return PTR_ERR(task);
874 list_add_tail(&task->list, &list);
877 fprintf(fp, "# %8s %8s %8s %s\n", "pid", "tid", "ppid", "comm");
879 list_for_each_entry(task, &list, list)
880 task__print_level(task, fp, 0);
886 static int __cmd_report(struct report *rep)
889 struct perf_session *session = rep->session;
891 struct perf_data *data = session->data;
893 signal(SIGINT, sig_handler);
896 ret = perf_session__cpu_bitmap(session, rep->cpu_list,
899 ui__error("failed to set cpu bitmap\n");
902 session->itrace_synth_opts->cpu_bitmap = rep->cpu_bitmap;
905 if (rep->show_threads) {
906 ret = perf_read_values_init(&rep->show_threads_values);
911 ret = report__setup_sample_type(rep);
913 /* report__setup_sample_type() already showed error message */
923 ret = perf_session__process_events(session);
925 ui__error("failed to process sample\n");
930 return stats_print(rep);
933 return tasks_print(rep, stdout);
935 report__warn_kptr_restrict(rep);
937 evlist__for_each_entry(session->evlist, pos)
938 rep->nr_entries += evsel__hists(pos)->nr_entries;
940 if (use_browser == 0) {
942 perf_session__fprintf(session, stdout);
945 perf_session__fprintf_dsos(session, stdout);
948 perf_session__fprintf_nr_events(session, stdout);
949 perf_evlist__fprintf_nr_events(session->evlist, stdout);
954 ret = report__collapse_hists(rep);
956 ui__error("failed to process hist entry\n");
964 * recalculate number of entries after collapsing since it
965 * might be changed during the collapse phase.
968 evlist__for_each_entry(session->evlist, pos)
969 rep->nr_entries += evsel__hists(pos)->nr_entries;
971 if (rep->nr_entries == 0) {
972 ui__error("The %s data has no samples!\n", data->path);
976 report__output_resort(rep);
978 if (rep->total_cycles_mode) {
979 int block_hpps[6] = {
980 PERF_HPP_REPORT__BLOCK_TOTAL_CYCLES_PCT,
981 PERF_HPP_REPORT__BLOCK_LBR_CYCLES,
982 PERF_HPP_REPORT__BLOCK_CYCLES_PCT,
983 PERF_HPP_REPORT__BLOCK_AVG_CYCLES,
984 PERF_HPP_REPORT__BLOCK_RANGE,
985 PERF_HPP_REPORT__BLOCK_DSO,
988 rep->block_reports = block_info__create_report(session->evlist,
991 &rep->nr_block_reports);
992 if (!rep->block_reports)
996 return report__browse_hists(rep);
1000 report_parse_callchain_opt(const struct option *opt, const char *arg, int unset)
1002 struct callchain_param *callchain = opt->value;
1004 callchain->enabled = !unset;
1009 symbol_conf.use_callchain = false;
1010 callchain->mode = CHAIN_NONE;
1014 return parse_callchain_report_opt(arg);
1018 parse_time_quantum(const struct option *opt, const char *arg,
1019 int unset __maybe_unused)
1021 unsigned long *time_q = opt->value;
1024 *time_q = strtoul(arg, &end, 0);
1028 pr_err("time quantum cannot be 0");
1031 end = skip_spaces(end);
1034 if (!strcmp(end, "s")) {
1035 *time_q *= NSEC_PER_SEC;
1038 if (!strcmp(end, "ms")) {
1039 *time_q *= NSEC_PER_MSEC;
1042 if (!strcmp(end, "us")) {
1043 *time_q *= NSEC_PER_USEC;
1046 if (!strcmp(end, "ns"))
1049 pr_err("Cannot parse time quantum `%s'\n", arg);
1054 report_parse_ignore_callees_opt(const struct option *opt __maybe_unused,
1055 const char *arg, int unset __maybe_unused)
1058 int err = regcomp(&ignore_callees_regex, arg, REG_EXTENDED);
1061 regerror(err, &ignore_callees_regex, buf, sizeof(buf));
1062 pr_err("Invalid --ignore-callees regex: %s\n%s", arg, buf);
1065 have_ignore_callees = 1;
1072 parse_branch_mode(const struct option *opt,
1073 const char *str __maybe_unused, int unset)
1075 int *branch_mode = opt->value;
1077 *branch_mode = !unset;
1082 parse_percent_limit(const struct option *opt, const char *str,
1083 int unset __maybe_unused)
1085 struct report *rep = opt->value;
1086 double pcnt = strtof(str, NULL);
1088 rep->min_percent = pcnt;
1089 callchain_param.min_percent = pcnt;
1093 int cmd_report(int argc, const char **argv)
1095 struct perf_session *session;
1096 struct itrace_synth_opts itrace_synth_opts = { .set = 0, };
1098 bool has_br_stack = false;
1099 int branch_mode = -1;
1101 bool branch_call_mode = false;
1102 #define CALLCHAIN_DEFAULT_OPT "graph,0.5,caller,function,percent"
1103 static const char report_callchain_help[] = "Display call graph (stack chain/backtrace):\n\n"
1104 CALLCHAIN_REPORT_HELP
1105 "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT;
1106 char callchain_default_opt[] = CALLCHAIN_DEFAULT_OPT;
1107 const char * const report_usage[] = {
1108 "perf report [<options>]",
1111 struct report report = {
1113 .sample = process_sample_event,
1114 .mmap = perf_event__process_mmap,
1115 .mmap2 = perf_event__process_mmap2,
1116 .comm = perf_event__process_comm,
1117 .namespaces = perf_event__process_namespaces,
1118 .cgroup = perf_event__process_cgroup,
1119 .exit = perf_event__process_exit,
1120 .fork = perf_event__process_fork,
1121 .lost = perf_event__process_lost,
1122 .read = process_read_event,
1123 .attr = perf_event__process_attr,
1124 .tracing_data = perf_event__process_tracing_data,
1125 .build_id = perf_event__process_build_id,
1126 .id_index = perf_event__process_id_index,
1127 .auxtrace_info = perf_event__process_auxtrace_info,
1128 .auxtrace = perf_event__process_auxtrace,
1129 .event_update = perf_event__process_event_update,
1130 .feature = process_feature_event,
1131 .ordered_events = true,
1132 .ordering_requires_timestamps = true,
1134 .max_stack = PERF_MAX_STACK_DEPTH,
1135 .pretty_printing_style = "normal",
1136 .socket_filter = -1,
1137 .annotation_opts = annotation__default_options,
1139 const struct option options[] = {
1140 OPT_STRING('i', "input", &input_name, "file",
1142 OPT_INCR('v', "verbose", &verbose,
1143 "be more verbose (show symbol address, etc)"),
1144 OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any message"),
1145 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1146 "dump raw trace in ASCII"),
1147 OPT_BOOLEAN(0, "stats", &report.stats_mode, "Display event stats"),
1148 OPT_BOOLEAN(0, "tasks", &report.tasks_mode, "Display recorded tasks"),
1149 OPT_BOOLEAN(0, "mmaps", &report.mmaps_mode, "Display recorded tasks memory maps"),
1150 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
1151 "file", "vmlinux pathname"),
1152 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
1153 "don't load vmlinux even if found"),
1154 OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1155 "file", "kallsyms pathname"),
1156 OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
1157 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1158 "load module symbols - WARNING: use only with -k and LIVE kernel"),
1159 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
1160 "Show a column with the number of samples"),
1161 OPT_BOOLEAN('T', "threads", &report.show_threads,
1162 "Show per-thread event counters"),
1163 OPT_STRING(0, "pretty", &report.pretty_printing_style, "key",
1164 "pretty printing style key: normal raw"),
1165 OPT_BOOLEAN(0, "tui", &report.use_tui, "Use the TUI interface"),
1166 OPT_BOOLEAN(0, "gtk", &report.use_gtk, "Use the GTK2 interface"),
1167 OPT_BOOLEAN(0, "stdio", &report.use_stdio,
1168 "Use the stdio interface"),
1169 OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
1170 OPT_BOOLEAN(0, "header-only", &report.header_only,
1171 "Show only data header."),
1172 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1173 sort_help("sort by key(s):")),
1174 OPT_STRING('F', "fields", &field_order, "key[,keys...]",
1175 sort_help("output field(s): overhead period sample ")),
1176 OPT_BOOLEAN(0, "show-cpu-utilization", &symbol_conf.show_cpu_utilization,
1177 "Show sample percentage for different cpu modes"),
1178 OPT_BOOLEAN_FLAG(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
1179 "Show sample percentage for different cpu modes", PARSE_OPT_HIDDEN),
1180 OPT_STRING('p', "parent", &parent_pattern, "regex",
1181 "regex filter to identify parent, see: '--sort parent'"),
1182 OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
1183 "Only display entries with parent-match"),
1184 OPT_CALLBACK_DEFAULT('g', "call-graph", &callchain_param,
1185 "print_type,threshold[,print_limit],order,sort_key[,branch],value",
1186 report_callchain_help, &report_parse_callchain_opt,
1187 callchain_default_opt),
1188 OPT_BOOLEAN(0, "children", &symbol_conf.cumulate_callchain,
1189 "Accumulate callchains of children and show total overhead as well. "
1190 "Enabled by default, use --no-children to disable."),
1191 OPT_INTEGER(0, "max-stack", &report.max_stack,
1192 "Set the maximum stack depth when parsing the callchain, "
1193 "anything beyond the specified depth will be ignored. "
1194 "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
1195 OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
1196 "alias for inverted call graph"),
1197 OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
1198 "ignore callees of these functions in call graphs",
1199 report_parse_ignore_callees_opt),
1200 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1201 "only consider symbols in these dsos"),
1202 OPT_STRING('c', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1203 "only consider symbols in these comms"),
1204 OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1205 "only consider symbols in these pids"),
1206 OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1207 "only consider symbols in these tids"),
1208 OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1209 "only consider these symbols"),
1210 OPT_STRING(0, "symbol-filter", &report.symbol_filter_str, "filter",
1211 "only show symbols that (partially) match with this filter"),
1212 OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
1214 "don't try to adjust column width, use these fixed values"),
1215 OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1216 "separator for columns, no spaces will be added between "
1217 "columns '.' is reserved."),
1218 OPT_BOOLEAN('U', "hide-unresolved", &symbol_conf.hide_unresolved,
1219 "Only display entries resolved to a symbol"),
1220 OPT_CALLBACK(0, "symfs", NULL, "directory",
1221 "Look for files with symbols relative to this directory",
1222 symbol__config_symfs),
1223 OPT_STRING('C', "cpu", &report.cpu_list, "cpu",
1224 "list of cpus to profile"),
1225 OPT_BOOLEAN('I', "show-info", &report.show_full_info,
1226 "Display extended information about perf.data file"),
1227 OPT_BOOLEAN(0, "source", &report.annotation_opts.annotate_src,
1228 "Interleave source code with assembly code (default)"),
1229 OPT_BOOLEAN(0, "asm-raw", &report.annotation_opts.show_asm_raw,
1230 "Display raw encoding of assembly instructions (default)"),
1231 OPT_STRING('M', "disassembler-style", &report.annotation_opts.disassembler_style, "disassembler style",
1232 "Specify disassembler style (e.g. -M intel for intel syntax)"),
1233 OPT_STRING(0, "prefix", &report.annotation_opts.prefix, "prefix",
1234 "Add prefix to source file path names in programs (with --prefix-strip)"),
1235 OPT_STRING(0, "prefix-strip", &report.annotation_opts.prefix_strip, "N",
1236 "Strip first N entries of source file path name in programs (with --prefix)"),
1237 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
1238 "Show a column with the sum of periods"),
1239 OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group, &report.group_set,
1240 "Show event group information together"),
1241 OPT_INTEGER(0, "group-sort-idx", &symbol_conf.group_sort_idx,
1242 "Sort the output by the event at the index n in group. "
1243 "If n is invalid, sort by the first event. "
1244 "WARNING: should be used on grouped events."),
1245 OPT_CALLBACK_NOOPT('b', "branch-stack", &branch_mode, "",
1246 "use branch records for per branch histogram filling",
1248 OPT_BOOLEAN(0, "branch-history", &branch_call_mode,
1249 "add last branch records to call history"),
1250 OPT_STRING(0, "objdump", &report.annotation_opts.objdump_path, "path",
1251 "objdump binary to use for disassembly and annotations"),
1252 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
1253 "Disable symbol demangling"),
1254 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
1255 "Enable kernel symbol demangling"),
1256 OPT_BOOLEAN(0, "mem-mode", &report.mem_mode, "mem access profile"),
1257 OPT_INTEGER(0, "samples", &symbol_conf.res_sample,
1258 "Number of samples to save per histogram entry for individual browsing"),
1259 OPT_CALLBACK(0, "percent-limit", &report, "percent",
1260 "Don't show entries under that percent", parse_percent_limit),
1261 OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1262 "how to display percentage of filtered entries", parse_filter_percentage),
1263 OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
1264 "Instruction Tracing options\n" ITRACE_HELP,
1265 itrace_parse_synth_opts),
1266 OPT_BOOLEAN(0, "full-source-path", &srcline_full_filename,
1267 "Show full source file name path for source lines"),
1268 OPT_BOOLEAN(0, "show-ref-call-graph", &symbol_conf.show_ref_callgraph,
1269 "Show callgraph from reference event"),
1270 OPT_BOOLEAN(0, "stitch-lbr", &report.stitch_lbr,
1271 "Enable LBR callgraph stitching approach"),
1272 OPT_INTEGER(0, "socket-filter", &report.socket_filter,
1273 "only show processor socket that match with this filter"),
1274 OPT_BOOLEAN(0, "raw-trace", &symbol_conf.raw_trace,
1275 "Show raw trace event output (do not use print fmt or plugins)"),
1276 OPT_BOOLEAN(0, "hierarchy", &symbol_conf.report_hierarchy,
1277 "Show entries in a hierarchy"),
1278 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
1279 "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
1280 stdio__config_color, "always"),
1281 OPT_STRING(0, "time", &report.time_str, "str",
1282 "Time span of interest (start,stop)"),
1283 OPT_BOOLEAN(0, "inline", &symbol_conf.inline_name,
1284 "Show inline function"),
1285 OPT_CALLBACK(0, "percent-type", &report.annotation_opts, "local-period",
1286 "Set percent type local/global-period/hits",
1287 annotate_parse_percent_type),
1288 OPT_BOOLEAN(0, "ns", &symbol_conf.nanosecs, "Show times in nanosecs"),
1289 OPT_CALLBACK(0, "time-quantum", &symbol_conf.time_quantum, "time (ms|us|ns|s)",
1290 "Set time quantum for time sort key (default 100ms)",
1291 parse_time_quantum),
1292 OPTS_EVSWITCH(&report.evswitch),
1293 OPT_BOOLEAN(0, "total-cycles", &report.total_cycles_mode,
1294 "Sort all blocks by 'Sampled Cycles%'"),
1297 struct perf_data data = {
1298 .mode = PERF_DATA_MODE_READ,
1300 int ret = hists__init();
1306 ret = perf_config(report__config, &report);
1310 argc = parse_options(argc, argv, options, report_usage, 0);
1313 * Special case: if there's an argument left then assume that
1314 * it's a symbol filter:
1317 usage_with_options(report_usage, options);
1319 report.symbol_filter_str = argv[0];
1322 if (annotate_check_args(&report.annotation_opts) < 0)
1325 if (report.mmaps_mode)
1326 report.tasks_mode = true;
1329 perf_quiet_option();
1331 if (symbol_conf.vmlinux_name &&
1332 access(symbol_conf.vmlinux_name, R_OK)) {
1333 pr_err("Invalid file: %s\n", symbol_conf.vmlinux_name);
1336 if (symbol_conf.kallsyms_name &&
1337 access(symbol_conf.kallsyms_name, R_OK)) {
1338 pr_err("Invalid file: %s\n", symbol_conf.kallsyms_name);
1342 if (report.inverted_callchain)
1343 callchain_param.order = ORDER_CALLER;
1344 if (symbol_conf.cumulate_callchain && !callchain_param.order_set)
1345 callchain_param.order = ORDER_CALLER;
1347 if ((itrace_synth_opts.callchain || itrace_synth_opts.add_callchain) &&
1348 (int)itrace_synth_opts.callchain_sz > report.max_stack)
1349 report.max_stack = itrace_synth_opts.callchain_sz;
1351 if (!input_name || !strlen(input_name)) {
1352 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
1355 input_name = "perf.data";
1358 data.path = input_name;
1359 data.force = symbol_conf.force;
1362 session = perf_session__new(&data, false, &report.tool);
1363 if (IS_ERR(session))
1364 return PTR_ERR(session);
1366 ret = evswitch__init(&report.evswitch, session->evlist, stderr);
1370 if (zstd_init(&(session->zstd_data), 0) < 0)
1371 pr_warning("Decompression initialization failed. Reported data may be incomplete.\n");
1373 if (report.queue_size) {
1374 ordered_events__set_alloc_size(&session->ordered_events,
1378 session->itrace_synth_opts = &itrace_synth_opts;
1380 report.session = session;
1382 has_br_stack = perf_header__has_feat(&session->header,
1383 HEADER_BRANCH_STACK);
1384 if (perf_evlist__combined_sample_type(session->evlist) & PERF_SAMPLE_STACK_USER)
1385 has_br_stack = false;
1387 setup_forced_leader(&report, session->evlist);
1389 if (symbol_conf.group_sort_idx && !session->evlist->nr_groups) {
1390 parse_options_usage(NULL, options, "group-sort-idx", 0);
1395 if (itrace_synth_opts.last_branch || itrace_synth_opts.add_last_branch)
1396 has_br_stack = true;
1398 if (has_br_stack && branch_call_mode)
1399 symbol_conf.show_branchflag_count = true;
1401 memset(&report.brtype_stat, 0, sizeof(struct branch_type_stat));
1404 * Branch mode is a tristate:
1405 * -1 means default, so decide based on the file having branch data.
1406 * 0/1 means the user chose a mode.
1408 if (((branch_mode == -1 && has_br_stack) || branch_mode == 1) &&
1409 !branch_call_mode) {
1410 sort__mode = SORT_MODE__BRANCH;
1411 symbol_conf.cumulate_callchain = false;
1413 if (branch_call_mode) {
1414 callchain_param.key = CCKEY_ADDRESS;
1415 callchain_param.branch_callstack = true;
1416 symbol_conf.use_callchain = true;
1417 callchain_register_param(&callchain_param);
1418 if (sort_order == NULL)
1419 sort_order = "srcline,symbol,dso";
1422 if (report.mem_mode) {
1423 if (sort__mode == SORT_MODE__BRANCH) {
1424 pr_err("branch and mem mode incompatible\n");
1427 sort__mode = SORT_MODE__MEMORY;
1428 symbol_conf.cumulate_callchain = false;
1431 if (symbol_conf.report_hierarchy) {
1432 /* disable incompatible options */
1433 symbol_conf.cumulate_callchain = false;
1436 pr_err("Error: --hierarchy and --fields options cannot be used together\n");
1437 parse_options_usage(report_usage, options, "F", 1);
1438 parse_options_usage(NULL, options, "hierarchy", 0);
1442 perf_hpp_list.need_collapse = true;
1445 if (report.use_stdio)
1447 else if (report.use_tui)
1449 else if (report.use_gtk)
1452 /* Force tty output for header output and per-thread stat. */
1453 if (report.header || report.header_only || report.show_threads)
1455 if (report.header || report.header_only)
1456 report.tool.show_feat_hdr = SHOW_FEAT_HEADER;
1457 if (report.show_full_info)
1458 report.tool.show_feat_hdr = SHOW_FEAT_HEADER_FULL_INFO;
1459 if (report.stats_mode || report.tasks_mode)
1461 if (report.stats_mode && report.tasks_mode) {
1462 pr_err("Error: --tasks and --mmaps can't be used together with --stats\n");
1466 if (report.total_cycles_mode) {
1467 if (sort__mode != SORT_MODE__BRANCH)
1468 report.total_cycles_mode = false;
1473 if (strcmp(input_name, "-") != 0)
1474 setup_browser(true);
1478 if (sort_order && strstr(sort_order, "ipc")) {
1479 parse_options_usage(report_usage, options, "s", 1);
1483 if (sort_order && strstr(sort_order, "symbol")) {
1484 if (sort__mode == SORT_MODE__BRANCH) {
1485 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1486 sort_order, "ipc_lbr");
1487 report.symbol_ipc = true;
1489 snprintf(sort_tmp, sizeof(sort_tmp), "%s,%s",
1490 sort_order, "ipc_null");
1493 sort_order = sort_tmp;
1496 if ((last_key != K_SWITCH_INPUT_DATA && last_key != K_RELOAD) &&
1497 (setup_sorting(session->evlist) < 0)) {
1499 parse_options_usage(report_usage, options, "s", 1);
1501 parse_options_usage(sort_order ? NULL : report_usage,
1506 if ((report.header || report.header_only) && !quiet) {
1507 perf_session__fprintf_info(session, stdout,
1508 report.show_full_info);
1509 if (report.header_only) {
1513 } else if (use_browser == 0 && !quiet &&
1514 !report.stats_mode && !report.tasks_mode) {
1515 fputs("# To display the perf.data header info, please use --header/--header-only options.\n#\n",
1520 * Only in the TUI browser we are doing integrated annotation,
1521 * so don't allocate extra space that won't be used in the stdio
1524 if (ui__has_annotation() || report.symbol_ipc ||
1525 report.total_cycles_mode) {
1526 ret = symbol__annotation_init();
1530 * For searching by name on the "Browse map details".
1531 * providing it only in verbose mode not to bloat too
1532 * much struct symbol.
1536 * XXX: Need to provide a less kludgy way to ask for
1537 * more space per symbol, the u32 is for the index on
1539 * See symbol__browser_index.
1541 symbol_conf.priv_size += sizeof(u32);
1542 symbol_conf.sort_by_name = true;
1544 annotation_config__init(&report.annotation_opts);
1547 if (symbol__init(&session->header.env) < 0)
1550 if (report.time_str) {
1551 ret = perf_time__parse_for_ranges(report.time_str, session,
1552 &report.ptime_range,
1558 itrace_synth_opts__set_time_range(&itrace_synth_opts,
1563 if (session->tevent.pevent &&
1564 tep_set_function_resolver(session->tevent.pevent,
1565 machine__resolve_kernel_addr,
1566 &session->machines.host) < 0) {
1567 pr_err("%s: failed to set libtraceevent function resolver\n",
1572 sort__setup_elide(stdout);
1574 ret = __cmd_report(&report);
1575 if (ret == K_SWITCH_INPUT_DATA || ret == K_RELOAD) {
1576 perf_session__delete(session);
1577 last_key = K_SWITCH_INPUT_DATA;
1583 if (report.ptime_range) {
1584 itrace_synth_opts__clear_time_range(&itrace_synth_opts);
1585 zfree(&report.ptime_range);
1588 if (report.block_reports) {
1589 block_info__free_report(report.block_reports,
1590 report.nr_block_reports);
1591 report.block_reports = NULL;
1594 zstd_fini(&(session->zstd_data));
1595 perf_session__delete(session);